PhD thesis of Tristan Petit (UBO) defended on 07/03/2017
Supervisors: Touria Bajjouk, Christophe Delacourt, Francis Gohin
Airborne hyperspectral imaging is a potential candidate for mapping and monitoring coral reefs at large scale and with high spatial resolution. In this thesis, we first present the processing steps to be applied to hyperspectral signals for extracting information about seabed types, bathymetry and water optical properties, and we discuss their efficiency with respect to two main confounding factors: (i) low signal to noise ratio of the measured signals, and (ii) large number and variability of physical interactions occurring between the entrance of sunlight into the atmosphere and its measurement by the hyperspectral sensor.
Considering these limitations, we examine the performance of an already existing water column processing method: semi-analytical model inversion by optimization. We first evaluate the robustness of seabed type and bathymetry estimation for six different inversion setups. The results on hyperspectral images acquired over Réunion Island reefs in 2009 show that the choice of the inversion setup plays an important role on the quality of the estimations and that the most widely used inversion setup does not always produce the best results.
We then evaluate the importance of the accuracy of the parameterization of the direct semi-analytical model. This is done through a sensitivity analysis performed on both simulated and real hyperspectral data acquired in Réunion Island in 2015. The analysis is performed for each inversion setup previously studied. This study shows that in coral reef context the accuracy of the parameterization of the direct model is less important than the choice of the inversionsetup. We also demonstrate that it is not possible to identify the most influent parameters of the direct model because it depends on the relative concentration of each optically active constituent.